---
title: 'Suno70k: AI Song Dataset for Cover Generation'
url: https://www.emergentmind.com/topics/suno70k
type: topic
---

# Suno70k: AI Song Dataset for Cover Generation

Suno70k is a large-scale AI song dataset introduced as part of the SongEcho framework for cover-song generation. It was constructed to address what the authors describe as a severe lack of open, high-quality, full-song data suitable for this task, particularly data containing full-length songs with lyrics, vocals, and accompaniment rather than short vocal clips or metadata-only records. In the SongEcho study, Suno70k functions both as a curated training corpus and as an evaluation substrate for melody-conditioned, text-conditioned full-song generation, while also being presented as a comparatively copyright-safer alternative because it relies exclusively on AI-generated music rather than human-recorded copyrighted songs [2602.19976].

## 1. Origin and intended problem setting

Suno70k is positioned within the specific problem of cover-song generation. In the underlying formulation, cover generation is treated as conditional generation that simultaneously produces new vocals and accompaniment conditioned on the original vocal melody and text prompts. The dataset is introduced because existing resources are described as unsuitable for one of three reasons: MIR-style datasets are small; large commercial-song corpora often provide metadata but no usable audio; and singing-voice datasets usually contain only short, single-track vocal clips rather than complete songs with vocals and accompaniment.

Within that framing, Suno70k is not a general-purpose music corpus. Its purpose is narrower and more technical: to provide full songs with lyrics and enriched tags in a form compatible with training and evaluating SongEcho, which is built on ACE-Step and conditioned on extracted vocal melody and text tags. A plausible implication is that the dataset’s design is tightly coupled to the control requirements of full-song generation rather than to canonical MIR tasks such as retrieval or transcription.

The source material is the publicly available “Suno.ai Music Generation dataset” hosted on Hugging Face. The paper states that the source collection contains 659,788 AI-generated songs with metadata and song links, but also notes that the quality is highly variable. Suno70k is therefore a curated derivative rather than a direct republication of the source corpus.

## 2. Curation pipeline and inclusion criteria

The construction of Suno70k follows a three-stage curation pipeline: filtering, quality assessment, and enhanced tagging [2602.19976].

The filtering stage removes entries with incomplete metadata such as missing IDs, missing lyrics, or missing tags; deduplicates by ID; excludes purely instrumental tracks; removes songs with unclear lyric structure, unrecognizable characters, or non-English lyrics; and enforces a duration cap of 4 minutes to match ACE-Step’s generation limit. These criteria define the corpus as English-language, lyric-bearing, full-song material within a bounded duration regime.

The quality-assessment stage downloads the audio and applies SongEval scoring on five dimensions: overall coherence, memorability, naturalness of vocal breathing and phrasing, clarity of song structure, and overall musicality. Any sample with a score below 3/5 in any one of these dimensions is discarded. This makes Suno70k not merely filtered for metadata completeness, but also aggressively filtered for estimated perceptual quality.

The enhanced-tagging stage addresses incomplete source metadata. Qwen2-audio is used to generate additional tags for genre, vocal type, instruments, and mood. These generated tags are concatenated with the original tags, deduplicated, and truncated to at most 20 tags per song, separated by commas in ACE-Step-compatible format. This suggests that the dataset is intended to support prompt-conditioned generation with a relatively standardized symbolic conditioning channel.

## 3. Scale, composition, and annotation structure

In scale, Suno70k contains 69,469 songs total, with 69,379 training songs and 90 test songs, for about 3,000 hours of audio [2602.19976]. The paper does not report an artist count, genre count, or explicit vocal/accompaniment track counts in the style of a conventional MIR dataset, but it emphasizes that the corpus consists of full songs rather than short vocal snippets.

The annotation structure is song-level and multimodal. Every included track has lyrics, original tags, and generated tags spanning genre, vocal type, instruments, and mood. The lyrics are described as native lyrics from the Suno source rather than post hoc transcriptions. The paper further states that they are formatted line-by-line in a way that preserves phrase boundaries.

That line-separated lyric formatting is presented as operationally important. The appendix contrasts Suno70k with SongEval by noting that SongEval’s transcribed lyrics often have punctuation and segmentation errors, whereas Suno70k’s native lyrics preserve phrase boundaries line-by-line. The authors argue that this structure implicitly helps lyric-melody alignment during training and reduces PER. A plausible implication is that Suno70k’s annotations are not only richer than raw tags, but structurally more aligned with sequence-conditioned generative modeling.

## 4. Role in SongEcho training and evaluation

Suno70k is the principal dataset on which SongEcho is trained and tested. SongEcho is built on ACE-Step, and the model is trained on Suno70k while conditioning on extracted vocal melody (\(F0\)) and text tags. The paper states that the experiments report quantitative results on the Suno70k test set and also use Suno70k as a reference dataset for aesthetics evaluation and tag-swap robustness tests.

The training objective reported for the model is a diffusion-style loss:
$$
\mathcal{L}_{\text{FM} = \mathbb{E}_{x_0, z \sim \mathcal{N}(0,I), t \sim U[0, 1]} \left[ \left\| \left( \epsilon_\theta (x_t, t_{tag}, l, t, p) \cdot (-\sigma_t) + x_t \right) - x_0 \right\|_2^2 \right],
$$
where the conditioning inputs include text tags \(t_{tag}\), lyrics \(l\), and pitch sequence \(p\). Suno70k provides the lyrics and tags used in this regime.

The melody-conditioning path depends on the dataset through lyric-aligned songs and extracted pitch features. The paper states that \(F0\) is extracted at 100 Hz using RMVPE and that voiced components are normalized between 50–900 Hz, then concatenated with a voiced/unvoiced flag to form the melody feature. The appendix further discusses a degeneracy in which full-audio conditioning can collapse to \(\gamma_m \approx \mathbf{0}\), \(\beta_m \approx m\), effectively copying the condition; for melody control, the paper argues that this shortcut is not acceptable because it would discard timbre and lyrics. Suno70k is the concrete training corpus in which this conditioning distinction is operationalized.

The dataset is also used in a special “swapped tags” evaluation, where text tags in the test set are randomly swapped to test whether melody control remains stable under mismatched text prompts. The authors interpret stability in this setting as evidence that the model prioritizes source melody over text when the two conflict.

## 5. Quantitative performance associated with the dataset

The strongest quantitative evidence reported for Suno70k is indirect: it enables SongEcho to be trained as a cover-song generator and evaluated against multiple baselines on a shared test set [2602.19976]. The comparison includes ACE-Step alone, ACE-Step+SA ControlNet, ACE-Step+SA ControlNet+LoRA, and ACE-Step+MuseControlLite.

On the Suno70k test set, SongEcho achieves the best overall scores reported in the paper: RPA 0.7080, RCA 0.7339, OA 0.6952, CLAP 0.3243, FD 42.06, KL 0.1123, and PER 0.2951. The paper states that these results are stronger than the baselines, particularly in melody control and distributional/audio quality metrics.

The study also highlights parameter efficiency. SongEcho uses 49.1M trainable parameters, which is reported as 3.07% of ACE-Step+SA ControlNet, 14.8% of ACE-Step+SA ControlNet+LoRA, and 26.0% of ACE-Step+MuseControlLite. An ablation table further shows that reducing training data from the full Suno70k to 1,000 samples still gives reasonably strong results, whereas 100 samples is clearly insufficient. This suggests that the dataset supports a regime in which both scale and data efficiency can be studied.

The appendix additionally reports SongEval aesthetic scores across methods on both Suno70k and SongEval. On Suno70k, SongEcho scores 3.776 coherence, 3.485 musicality, 3.644 memorability, 3.534 clarity, and 3.440 naturalness. These values are used to support the claim that the model performs best on all five aesthetic dimensions in that comparison.

## 6. Relation to other datasets and resource landscape

The paper situates Suno70k against several dataset families that are said to be inadequate for cover-song generation [2602.19976]. MIR-1K, MIR-ST500, and Cmedia are described as having rich annotations but being too small. WASABI, Million Song Dataset, and SongCompose-PT are described as large but mostly metadata without raw audio. OpenSinger, M4Singer, and GTSinger are characterized as useful for singing synthesis but too short and single-track for cover-song generation.

SongEval occupies a more closely related position. It is treated as full-song AI music, but with inconsistent quality and originally missing lyrics and tags. In a separate comparison, the authors augment SongEval by generating tags with Qwen2-audio and lyrics with Whisper plus All-in-One, then select the top 100 English songs by aesthetic score and remove six that ACE-Step cannot parse, leaving 94 songs for evaluation. This comparison matters because it explicitly contrasts Suno70k’s native lyrics and annotations with the noisier transcription-based annotations used for SongEval.

In that comparison, the authors argue that Suno70k’s native line-separated lyrics are advantageous for phrase-level alignment, while transcription errors in SongEval degrade alignment. A plausible implication is that Suno70k is not simply larger or cleaner than nearby alternatives, but structurally better matched to lyric-conditioned and melody-conditioned generation.

## 7. Limitations, bias, and ethical framing

The paper’s ethical framing is explicit: training was done exclusively on AI-generated music, and the outputs are intended for non-commercial academic demonstration only [2602.19976]. The use of AI-generated source material is presented as a way to mitigate copyright problems associated with human-recorded songs.

At the same time, the paper identifies several limitations and biases. Suno70k is filtered to English-language lyrics, making it language-biased and excluding non-English song content. It excludes instrumental tracks and songs over 4 minutes, favoring a specific full-song format rather than the broader music spectrum. The SongEval-based quality filter introduces an aesthetic bias because only songs that score well on that automated evaluator are retained. The tags are generated or augmented by Qwen2-audio, which can introduce annotation noise or systematic bias.

The paper also notes a representational limitation arising from the source domain itself. Because the source material is AI-generated rather than human-performed, Suno70k lacks the subtle expressive nuances and fine-grained vocal technique annotations of professional singing. The authors state that this limits the realism of cover-song generation and especially micro-level performance control.

Taken together, these constraints define Suno70k as a task-specific, highly curated corpus rather than a neutral sample of musical practice. Its significance lies less in universal coverage than in providing an open, filtered, annotated, full-song dataset tailored to the technical requirements of cover-song generation under melody and text conditioning.

Source: https://www.emergentmind.com/topics/suno70k